• Login
  • Register

Work for a Member company and need a Member Portal account? Register here with your company email address.

Project

Graph RAG as Human Choice Model: Building a Data-Driven Simulation Framework with LLM Agents

The study of human behavior in urban environment is an important field in city sciences. However, it's usually not easy to collect behavior data especially in newly developed area. Recently, generative agents provide a way to simulate human behaviors without data using llms. But these generative methods still face the challenge of reliability, as they may not always produce consistent or realistic behavior simulations.

To address this, a hybrid approach is proposed, combining Graph Retrieval-Augmented Generation (Graph RAG) with LLMs. By integrating graph-based retrieval, the model aims to enhance the contextual understanding and accuracy of the generated behaviors with a small amount of data, making the simulations more reliable and reflective of real-world dynamics. 

Framework

Great thanks to Terasa's amazing researches, the original data is from her work on amenities distribution (https://github.com/CityScope/AmenitiesDistribution.git). It is the trip and demongraphic data in Boston from replica (https://documentation.replicahq.com/docs/transit), combined with amenity analysis.

BDI Behavior Graph

The construction of behavior graph in this model follows the schema of BDI agents (https://gama-platform.org/wiki/BDIAgents).

  • Belief: What does the agent know. (profile)
  • Desire: What the agent want to do. ('Eat','Shop' or 'Recreation')
  • Intension: The final choice the agent made. (amenity choice, mobility choice)

Graph recommondation